OpenAI touts big gains in COBOL migration with GPT-6 models

OpenAI announced on Monday that its latest models, including GPT-6 Astra and GPT-5.6 Sol, show substantial improvements in migrating legacy programming languages to newer ones, according to gizmodo.com.

The company introduced two new benchmarks to measure the work. MultiLangBench tests whether a model can build SQLite from scratch in a given language, while a second test checks whether a model can migrate COBOL programs to C++, Java, Kotlin, Python or Rust without changing their behavior.

OpenAI said GPT-6 scored the highest on MultiLangBench with an average pass rate of about 87%. On the COBOL migration test, GPT-6.1 Sol with high reasoning effort came close to 100%, according to the company.

OpenAI said the research was guided by feedback from developers and enterprise customers. The company did not provide a timeline for when the benchmark results would be reflected in products.

Why COBOL still matters

COBOL was invented in 1959 and still handles 95% of all ATM transactions, giving companies a strong incentive to find faster ways to modernize old code. Earlier this year, Anthropic wrote in a blog post that its Claude Code tool could help businesses modernize COBOL codebases in quarters instead of years; Bloomberg reported the same day that IBM shares fell 13% after the post, their biggest single-day slide since 2000.